1. Field of the Invention
The present invention relates to an object estimation device, an object estimation method and an object estimation program for estimating position and size of an object in an image.
2. Description of the Related Art
As a method for detecting a specific object, such as a face or a person in an image, detection methods using an image pyramid have been proposed (see, for example, U.S. Patent Application Publication Nos. 2009324087 and 2008247651, which are hereinafter referred to as Patent Documents 1 and 2). In the techniques disclosed in Patent Documents 1 and 2, a plurality of smoothed images of different scales are generated by repeating convolution of an image with a Gaussian filter. Then, a differential image between each pair of the smoothed images of adjacent scales is generated (which is also referred to as a DOG (difference of Gaussian) image). The generation of the DOG images is carried out for each of multi-resolution images of different resolutions. Then, the maximum value and the minimum value in each DOG image are detected to detect the position of an object and to detect the size of the object in the multi-resolution images.
Further, a method for estimating an object area with using a Gaussian filter tailored to characteristics of the object has been proposed (see, for example, Japanese Unexamined Patent Publication No. 2003-248824, which is hereinafter referred to as Patent Document 3). Patent Document 3 discloses carrying out filtering with using a Gaussian filter, which has filtering characteristics tailored to the contour of an object to be detected, and then, highlighting an area having the contour of interest and diminishing the other areas, thereby detecting the object.
However, the techniques disclosed in Patent Documents 1 and 2 require generation of the differential image between each pair of smoothed images of adjacent scales, and it is necessary to generate a large number of differential images for a large number of resolution images. Therefore, speeding up of the operation is difficult to be achieved. Further, in the case where only the Gaussian filter based on the contour is applied, as disclosed in Patent Document 3, it is impossible to achieve accurate object detection.